pymss-ara
https://github.com/pymss-project/pymss-ara
It also doesn't work in real time, but it allows separations inside your DAW premade in a separate process (currently Windows and CUDA supported). The forte is, most current Roformers are supported.
demucs-rs
https://github.com/nikhilunni/demucs-rs
It has VST3 CLAP variant working on MacOS using Metal acceleration. Or the regular CLI variant with Vulkan backend on Windows/Linux (faster than DirectML in UVR)